MR Fingerprinting & Quantitative Imaging

Monday 1 June 2015

Debra McGivney¹, Anagha Deshmane², Yun Jiang², Dan Ma², and Mark Griswold^1,2
1Radiology, Case Western Reserve University, Cleveland, Ohio, United States, ²Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, United States

Magnetic resonance fingerprinting (MRF) is a technique that can provide quantitative maps of tissue parameters such as T1 and T2 relaxation times through matching observed signals to a precomputed complex-valued dictionary of modeled signal evolutions. Since each dictionary entry is uniquely defined by two real parameters, specifically T1 and T2, we propose to compress the dictionary onto a real-valued manifold of three dimensions using the nonlinear dimensionality reduction technique of kernel principal component analysis. Once the compression is achieved, we explore new computational applications for MRF, namely solving the partial volume problem.

Debra McGivney¹, Anagha Deshmane², Yun Jiang², Dan Ma², and Mark Griswold^1,2
1Radiology, Case Western Reserve University, Cleveland, Ohio, United States, ²Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, United States

Magnetic Resonance Fingerprinting (MRF) can produce quantitative maps of tissue parameters such as T1 and T2 relaxation times by matching acquired signals to a predefined dictionary of signal evolutions. One inherent issue is that all voxels are assigned only one dictionary entry, even if they exhibit the partial volume effect. We apply a Bayesian statistical framework to solve the general partial volume problem for MRF without assigning in advance the specific dictionary entries that comprise a signal from one of these mixed voxels, rather, assumptions are made on the probability distributions of the mixed signals and their component signals.

Philippe Pouliot^1,2, Louis Gagnon³, Tina Lam⁴, Pramod Avti⁵, Michèle Desjardins¹, Ashok Kakkar⁴, Sava Sakadzic³, David Boas³, and Frédéric Lesage^1
1Electrical Engineering, Ecole Polytechnique Montreal, Montreal, QC, Canada, ²Research Centre, Montreal Heart Institute, Montreal, QC, Canada, ³Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, MA, United States, ⁴Chemistry Department, McGill University, QC, Canada, ⁵Montreal Heart Institute, QC, Canada

MR vascular fingerprinting is a novel approach to estimate cerebral blood volume, vessel radius and oxygenation. To our knowledge, this approach has not yet been fully validated. Here we implemented the sequence in mice and exploited a dictionary built on simulations of the MR signal based on realistic vasculature built on 2-photon angiograms. A dictionary for fingerprint extraction was generated by sampling along 5 parameters: hemoglobin saturation, vessel radius, capillary density, SPION concentration and magnetic field inhomogeneity. Following linearization, the dictionary eigensystem was characterized. This confirmed that all its eigenvalues are positive and distinct, and therefore all parameters studied are theoretically identifiable.

Cristoffer Cordes¹ and Matthias Günther^1,2
1Fraunhofer MEVIS, Bremen, Germany, ²MR-Imaging and Spectroscopy, University of Bremen, Bremen, Germany

In order to extract the information density of images acquired with a given protocol, data was parameter mapped (T1, T2, M0) using an objective function based on a simulated signal model, minimized with a variation of the simulated annealing algorithm. Calculating the uncertainty volumes based on an uncertainty condition of the objective function reveals a contrast that is able to rank the performance of the utilized sequences by eliminating the sequence of least preferable impact in a greedy fashion. It also reveals the voxel-wise shape of the remaining flaws. The algorithm was tested on a series of TSE acquisitions.

Taejoon Eo¹, Jinseong Jang², Minoh Kim², Dong-hyun Kim², and Dosik Hwang^2
1Yonsei University, Seoul, Seoul, Korea, ²Yonsei University, Seoul, Korea

Proposed tier-specific WEST method could sufficiently suppress the noise-like artifacts in the maps obtained by the conventional WEST. Consequently, this method enables acquisition of accurate maps from extremely undersampled Cartesian MRF data.

Mathieu Sarracanie^1,2, Ouri Cohen¹, and Matthew S Rosen^1,2
1MGH/A.A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ²Department of Physics, Harvard University, Cambridge, MA, United States

2D MR Fingerprinting has recently been shown at low magnetic field. Here, we demonstrate MRF in 3D at 6.5 mT, using an optimized set of 15 flip angles and repetition times (FA/TR), in a Cartesian acquisition of k-space with a new hybrid b-SSFP-EPI sequence. We measure quantitative parameters in 3D, and generate several image contrasts in a single acquisition (proton density, T1, T2) in less than 30 minutes. The combination of 3D MRF with low field MRI scanners has great potential to provide clinically relevant contrast with portable low cost MR scanners.

Jesse Ian Hamilton¹, Katherine L Wright¹, Yun Jiang¹, Luis Hernandez-Garcia², Dan Ma¹, Mark Griswold^1,3, and Nicole Seiberlich^1,3
1Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, ²Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States, ³Radiology, Case Western Reserve University, Cleveland, OH, United States

A flexible framework for MR Fingerprinting pulse sequence design is presented that includes the MRI signal encoding, gridding, and pattern recognition directly in the optimization. The method was validated in a phantom study by designing sequence for mapping T1, T2, and M0 in under 3s using a highly undersampled spiral trajectory. Parameter maps obtained with the optimized sequence have fewer artifacts and higher agreement with spin echo measurements compared to unoptimized sequences. The optimization framework is easily generalizable to other MRF applications.

Christian Anderson¹, Ying Gao¹, Chris Flask^1,2, and Lan Lu^2,3
1Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, United States, ²Radiology, Case Western Reserve University, Cleveland, Ohio, United States, ³Urology, Case Western Reserve University, Cleveland, Ohio, United States

Magnetic resonance fingerprinting (MRF) offers rapid simultaneous multi-parametric quantification, and also provides the potential to generate maps of other parameters. We have developed a novel scheme named "Multi-Preparation MRF" (MP-MRF) that implements adaptable magnetization preparations periodically during the dynamic MRF acquisition. Our initial simulations of the MP-MRF methodology show sensitivity to diffusion and perfusion contrast and reasonable estimates of T1, T2, and velocity in Shepp-Logan phantoms.

Te-Ming Lin¹, Su-Chin Chiu¹, Cheng-Chieh Cheng¹, Wen-Chau Wu^1,2, and Hsiao-Wen Chung^1
1Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan, ²Graduate Institute of Oncology, National Taiwan University, Taipei, Taiwan

The signal acquisition number is related to the computational complexity during signal analysis in MR fingerprinting. In this study, we develop a contour area index and demonstrate a quantitative method to evaluate the mapping precision under different signal acquisition numbers. It has potential in evaluating different RF excitation schemes in MR fingerprinting.

Nicolas Pannetier^1,2 and Norbert Schuff^1,2
1Radiology, UCSF, San Francisco, California, United States, ²VAMC, San Francisco, CA, United States

We evaluate the use of kd-tree (a space partitioning data structure) to speed-up the matching process in magnetic resonance fingerprinting. We found that, in combination with PCA reduction, the matching time can be reduced by 2 to 3 order of magnitude while preserving the accuracy. The matching time, however, increases with noise level and the PCA threshold remains a key element to tune to achieve the best performance.

Dan Ma¹, Eric Y Pierre¹, Yun Jiang¹, Kawin Setsompop², Vikas Gulani³, and Mark A Griswold^3
1Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, ²A.A Martinos Center for Biomedical Engineering, MGH, Harvard Medical School, Boston, MA, United States, ³Radiology, Case Western Reserve University, Cleveland, OH, United States

The purpose of this study is to extend the 2D MR Fingerprinting (MRF) and MRF-Music framework to 3D acquisitions. Both methods were originally implemented in 2D acquisitions and have shown high scan efficiency for quantifying multiple tissue properties simultaneously. In addition to the multi-parameter quantification in MRF, the MRF-Music sequence was proposed to provide musical sounds that can dramatically improve the patients’ experience in the MR scanner. In this study, the MRF and MRF-Music sequences were implemented to achieve 3D coverage while still maintaining a high scan efficiency and providing desirable sounds. T1 and T2 values from phantom studies of the 3D slab selective MRF and MRF-Music methods showed good agreement to the values from the standard measurements. The T1, T2, off-resonance and M0 maps from 3D non-selective MRF and MRF-Music also showed promising results of achieving 3D isotropic quantitative mapping.

Florian Knoll^1,2, Martijn A Cloos^1,2, Thomas Koesters^1,2, Michael Zenge³, Ricardo Otazo^1,2, and Daniel K Sodickson^1,2
1Center for Advanced Imaging Innovation and Research (CAI2R), NYU School of Medicine, New York, NY, United States, ²Bernard and Irene Schwartz Center for Biomedical Imaging, Department of Radiology, NYU School of Medicine, New York, NY, United States, ³Siemens Medical Solutions USA, Malvern, PA, United States

Despite the extensive opportunities offered by PET-MR systems, their use is still far from routine clinical practice. While it is feasible to acquire PET data in about 5 minutes, collecting the clinically relevant variety of traditional MR contrasts requires substantially more time. This bottleneck formed by the traditional MR paradigm leads to inefficient use of the PET component. This work proposes a one-step procedure that merges the MR fingerprinting framework with the PET acquisition, and employs a dedicated multi-modality reconstruction to enable a 6 minute comprehensive PET-MR exam, which can provide the majority of clinical MR contrasts alongside quantitative parametric maps of the relaxation parameters (T1,T2) together with improved PET images.

Bernhard Strasser¹, Wolfgang Bogner¹, Peter Bär¹, Gilbert Hangel¹, Elisabeth Springer¹, Vlado Mlynarik¹, Mark A Griswold^2,3, Dan Ma², Yun Jiang², Mathias Nittka⁴, Haris Saybasili⁴, and Siegfried Trattnig^1
1MRCE, Department of Biomedical Imaging and Image-guided Therapy, University of Vienna, Vienna, Vienna, Austria, ²Department of Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, United States, ³Radiology, Case Western Reserve University, Cleveland, Ohio, United States, ⁴Siemens Healthcare USA, Inc., Chicago, Illinois, United States

Previously, MR Fingerprinting (MRF) has been presented as a new method for simultaneous quantitative mapping of different physical MR properties. In this study, the T1 and T2 values of MRF were compared to conventional T1- and T2-mapping methods in the brains of five volunteers at 1.5T. Each volunteer was measured five times with a TrueFISP and a FISP based spiral MRF sequence, an MP2RAGE and a multi echo spin echo sequence for conventional T1 and T2 maps, respectively. Both MRF sequences showed a similar reproducibility but seemed to slightly underestimate the T2-values in comparison to the conventional sequences.

Zhaohuan Zhang^1,2, Zhe Wang^2,3, Subashini Srinivasan^2,3, Kyunghyun Sung^2,3, and Daniel B. Ennis^2,3
1Department of Physics & Astronomy, Shanghai Jiao Tong University, Shanghai, China, ²Department of Radiological Sciences, University of California, Los Angles, CA, United States, ³Department of Bioengineering, University of California, Los Angles, CA, United States

The objective of this study was to evaluate the accuracy and precision of pseudorandom inversion recovery balanced steady-state free precession magnetic resonance fingerprinting (MRF) relaxometry (T1 and T2) estimates over a range of SNRs and the number of acquired TRs (NTR) using Bloch equation simulations. Under the condition of perfect sampling, the Bloch simulations defined a lower-bound acquisition requirement of SNR¡Ý5 and NTR¡Ý400 for accurate and precise T1 and T2 estimates when using MRF. This work also concluded that MRF provides nearly equivalent T1 and T2 estimates.

Thomas Amthor¹, Mariya Doneva¹, Peter Koken¹, Jochen Keupp¹, and Peter Börnert^1
1Philips Research Europe, Hamburg, Germany

We present a comparison of different pattern matching algorithms for tissue characterization based on Magnetic Resonance Fingerprinting. The applicability of a simple dot product approach and a number of machine learning algorithms is investigated for different parameter regimes. We find that, in many cases, machine learning algorithms can offer higher accuracy and faster matching.

Mariya Doneva¹, Thomas Amthor¹, Peter Koken¹, Jochen Keupp¹, and Peter Börnert^1
1Philips Research Europe, Hamburg, Germany

In this work we demonstrate a comprehensive accuracy analysis exemplified on a bSSFP-based MRF sequence, which allows predicting the accuracy of MRF in different parameter ranges and defining confidence areas for the performance of MRF.

Enhao Gong¹, Qiyuan Tian¹, John M Pauly¹, and Jennifer A McNab^2
1Electrical Engineering, STANFORD UNIVERSITY, Stanford, California, United States, ²Radiology, STANFORD UNIVERSITY, Stanford, California, United States

Low Signal-to-Noise Ratio (SNR), especially at high b-values, is a critical problem for Diffusion MRI (dMRI). Methods with different signal models may fail to reconstruct under-sampled data from noisy measurement. Diffusion MRI signal contains redundancy as a multi-dimensional signal in both k-space and q-space. Here we proposed a novel approach to recover signal without explicitly enforcing any physical signal model. The method is model-free but learns the multi-dimensional redundancy, including the redundancy between neighborhood voxels, different directions and low\high b-values, from training samples. A Dictionary Learning approach is used to recover under-sampled signals in q-space. Quantitative results demonstrate the method can more accurately predict high b-value signal (>3000s/mm2) from low b-value signal. Also it produces more accurate physiological metrics such as Generalized Fractional Anisotropy (GFA) and Orientation Distribution Function (ODF) that potentially help to resolve intra-voxel crossing fibers.

Andrea Vargas¹, Laurent Milot², Simon Graham¹, and Philip Beatty^1
1Medical Biophysics, University of Toronto, Toronto, Ontario, Canada, ²Sunnybrook Research Institute, Toronto, Canada

The use of variable flip angles for 3D fast spin echo sequences (3DFSE) have shown to alter contrast in T2-weighted images relative to conventional 2DFSE. While these alterations of contrast may be minimal in brain tissues, they can have a great consequences in body applications that encompass a wide range of T2 values. In this study we evaluate the performance of current methods that aim to correct T2-contrast in a cervix cancer application which has a wide range of T2 values (35 ms < T2 < 84 ms). We show that the differentiation between recurrent tumor and radiation fibrosis may be ambiguous at clinical echo times using 3DFSE.

Lili He¹, Jinghua Wang², Mark Smith³, and Nehal A. Parikh^1,4
1Center for Perinatal Research, The Research Institute at Nationwide Children's Hospital, Columbus, Ohio, United States, ²Center for Cognitive and Behavioral Brain Imaging, The Ohio State University, Columbus, Ohio, United States, ³Radiology Department, Nationwide Children's Hospital, Columbus, Ohio, United States, ⁴Department of Pediatrics, The Ohio State University College of Medicine, Columbus, Ohio, United States

Three-dimensional T1-weighted sequences such as MP-RAGE are extremely valuable to evaluate neonatal and infant brain injury/development. Yet, the lack of complete myelination and smaller head size results in comparatively lower quality images as compared to adult brains. In this study, we consider WM-GM contrast efficiency as an objective function to optimize neonatal MP-RAGE parameters under optimal k-space sampling by means of computer simulation. Quantitative analysis indicated that WM-GM contrast to noise efficiency of images acquired with our optimal parameters was 20% higher than those using parameters recommended by a published protocol; similarly, mean SNR efficiency was increased by approximately 150%.

Jonathan I. Tamir¹, Weitian Chen², Peng Lai², Martin Uecker¹, Shreyas S. Vasanawala³, and Michael Lustig^1
1Electrical Engineering and Computer Sciences, University of California, Berkeley, Berkeley, CA, United States, ²Global Applied Science Laboratory, GE Healthcare, Menlo Park, CA, United States, ³Radiology, Stanford University, Stanford, CA, United States

Fast Spin Echo (FSE) is widely used in MR imaging due to its speed and robustness to image artifacts. However, blurring due to T2 decay inhibits its use for 3D musculoskeletal imaging. By compensating for signal decay and reconstructing a time series of images, the blurring can be reduced. In this work we resample and reorder phase encodes over a longer echo train length to improve scan efficiency. We add a locally low rank constraint to improve the conditioning of the reconstruction, producing multicontrast 3D FSE images at clinically feasible scan times.

Yi-Cheng Hsu¹, Ying-Hua Chu¹, Shang-Yueh Tsai², Wen-Jui Kuo³, and Fa-Hsuan Lin^1
1Institute of Biomedical Engineering, National Taiwan University, Taipei, Taiwan, ²Institute of Applied Physic, National Chengchi University, Taipei, Taiwan,³Institute of Neuroscience, National Yang Ming University, Taipei, Taiwan

A MP trueFISP sequence for brain structural imaging was implemented and tested. Compared with MP RAGE using the same acquisition time, it improves the contrast from 40% to 80% with 37.8% noise increase due to a wider readout bandwidth.

Bing Wu¹, Nan Hong², and Zhenyu Zhou^1
1GE healthcare China, Beijing, Beijing, China, ²Peking university people's hospital, Beijing, China

T1 rho acquisition is often constrained to single slice due to the long TSL needed, which makes the cross-examination with other measurements such as resting state fMRI difficult. In this work, we develop a rapid T1-rho mapping method that utilizes single-shot EPI acquisition and multi-band excitation that completes a 2mm isotropic whole brain T1 rho mapping within 5 minutes, which allows this acquisition to be added in a Parkinson disease related clinical study.

Won-Joon Do¹, Seung Hong Choi², Eung Yeop Kim³, and Sung-Hong Park^1
1Korea Advanced Institute of Science and Technology, Daejeon, Korea, ²Department of Radiology, Seoul National University College of Medicine, Seoul, Korea,³Department of Radiology, Gachon University Gil Medical Center, Incheon, Korea

Dual-echo sequence allows us to acquire 3D T1 and T2*-weighted images simultaneously. The conflicting parameter conditions of T1and T2* contrasts can be resolved by echo-specific k-space reordering schemes. However, abrupt changes in scan conditions for the echo-specific k-space reordering can cause ringing artifacts. In this study, we propose a new approach of smooth transition in the regions of abrupt changes, to suppress the artifacts. The ringing artifacts in the echo-specific k-space reordered dual-echo sequence without the smooth transition could be effectively suppressed with the proposed approach and thus the image qualities became closer to those acquired with conventional single-echo sequences.

Galen D Reed¹, Reeve Ingle¹, Ken O Johnson¹, Juan M Santos¹, Bob S Hu², and William R Overall^1
1Heartvista, Menlo Park, California, United States, ²Cardiology, Palo Alto Medical Foundation, Menlo Park, California, United States

High resolution inversion recovery imaging of myocardium within small breath hold durations is challenging due to the need for segmented acquisitions and short readout windows. By combining the efficiency of parallel spiral imaging with a 2-dimensional field-of-view reduction, we designed a sequence that acquires 1.7 mm in-plane resolution images in a 7 heartbeat breath hold. The short acquisition window enabled repeating the sequence to obtain a series of images with different inversion times. The efficacy of multiple TI imaging with and without 2D outer volume suppression was demonstrated.

Reconstruction & Processing Algorithms

Monday 1 June 2015

Paniz Adipour¹ and Michael R. Smith^1,2
1Electrical and Computer Engineering, University of Calgary, Calgary, Alberta, Canada, ²Radiology, University of Calgary, Calgary, Alberta, Canada

Truncation artifacts appear in DFT reconstructions through discontinuities across the ends of the data set which mathematically is cyclic in k-space. A suggestion indicates that similar position dependent distortions will be present in CS reconstructions which repeatedly use the DFT. A comparison is made between standard Wavelet and Contourlet CS reconstructions and proposed high k-space extrapolation enabled (Hi-KEE) variants of these approaches. The CS-Contourlet outperforms the common CS-Wavelet in providing a better sparse representation of contour-shaped objects and detailed textures. The Hi-KEE-CS-Contourlet is shown to outperform the CS-Contourlet by providing a better position independent resolution solution.

Edward Li¹, Mohammad Javad Shafiee¹, Audrey Chung¹, Farzad Khalvati², Alexander Wong¹, and Masoom A Haider^3
1Systems Design Engineering, University of Waterloo, Waterloo, Ontario, Canada, ²Department of Medical Imaging, University of Toronto, Toronto, Ontario, Canada, ³Sunnybrook Health Sciences Center, Toronto, Ontario, Canada

Compressive sensing reduces MRI acquisition times but requires advanced sparse reconstruction algorithm to produce high-quality MR images. We propose a novel sparse reconstruction method using a cross-domain stochastically fully-connected random field (CD-SFCRF) for improved reconstruction from compressive sensing MRI data. Peak-to-peak signal-to-noise ratio (PSNR) analysis of CD-SFCRF and other methods using a prostate training phantom demonstrate that CD-SFCRF has the highest PSNR across all under-sampling ratios of radial MRI acquisitions. A visual comparison using real patient cases illustrate that CD-SFCRF can improve fine tissue detail and contrast preservation while eliminating under-sampling artifacts.

Michael R. Smith^1,2, Jordan Woehr¹, Mathew E. MacDonald^2,3, and Paniz Adipour^1
1Electrical and Computer Engineering, University of Calgary, Calgary, Alberta, Canada, ²Radiology, University of Calgary, Calgary, Alberta, Canada, ³Seaman MR Family Research Centre, University of Calgary, Calgary, Alberta, Canada

Truncated k-space data sets provide higher temporal resolution but compromise spatial resolution during DFT reconstruction. Compressed sensing, using under-sampled data, is used to improve spatial resolution while retaining temporal resolution. Certain Fourier domain properties can produce MRI CS reconstruction with resolutions that are dependent on the position of an object in the final reconstructed image. We demonstrate this position dependent resolution and propose two approaches to overcome it: Fourier Shift (FS) and Area Specific Additional Truncation (ASAT) image resolution enhancement pre-processing techniques.

Satoshi Ito¹, Mone Shibuya¹, Kenji Ito¹, and Yoshifumi Yamada^1
1Utsunomiya University, Utsunomiya, Tochigi, Japan

It is difficults to apply CS to images with rapid spatial phase variations, since not only the magnitude but also phase regularization is required in the CS framework. An iterative MRI reconstruction with separate magnitude and phase regularization was proposed for applications where magnitude and phase maps are both of interest. Since this method requires the approximation of phase regularizer to cope with phase unwrapping problem, it is roughly 10 times slower than conventional CS and the convergence is not guaranteed. In this article we propose a novel image reconstruction scheme for CS-MRI in which phase regularizer or symmetrical sampling trajectory are not required in the rather standard CS reconstruction scheme, but highly robust to rapid phase changes. The proposed method uses multi-frame complex transforms to introduce sparseness for the complex image data.

Matthias Weigel^1
1Radiological Physics, Dept. of Radiology and Nuclear Medicine, University of Basel Hospital, Basel, Switzerland

The extended phase graph (EPG) concept is a favorite approach for the rapid quantitation of magnetization response. However, users frequently have problems to properly program the framework. One major reason may be that care has to be taken with the complex Fourier domains of the transverse magnetization and their inherent symmetry relations. The present educational abstract depicts these issues and shows how RF pulses and gradients act differently on the magnetization components. Solutions to overcome the described issues are presented and discussed. Additionally, the author provides representative EPG software demonstrating the solutions.

Pavan Poojar¹, Bikkemane Jayadev Nutandev¹, Amaresha Sridhar Konar¹, Rashmi R Rao¹, Ramesh Venkatesan², and Sairam Geethanath^1
1Medical Imaging Research Centre, Dayananda Sagar Institutions, Bangalore, Karnataka, India, ²Wipro-GE Healthcare, Bangalore, Karnataka, India

Cardiac MRI scans demands rapid acquisition of images to avoid motion artifacts. Region of interest (ROI) selected will be sparse and leads to arbitrary k-space shape. Active contour in combination with convex optimization leads to new ROI based acquisition strategy which gives arbitrary k-space trajectories and optimized gradients based on the constraints for given ROI. Retrospective studies were carried out on six cardiac datasets for different accelerations (3x, 4x, 5x and 10x) and Normalized Root Mean Square Error was calculated. Future work includes reconstruction of image using ROI Compressed Sensing.

Weiyi Chen¹, Yi Guo¹, Ziyue Wu², and Krishna S. Nayak^1,2
1Electrical Engineering, University of Southern California, Los Angeles, CA, United States, ²Biomedical Engineering, University of Southern California, Los Angeles, CA, United States

We propose a method for MRI constrained reconstruction using ADMM framework that is data-driven, and does not require manual selection of tuning parameters. We use Morozov's discrepancy principle as a criterion to iteratively determine the tuning parameter. Tests with T2w brain data show that the reconstruction quality is comparable with reconstructions using manually selected parameter.

Xiaobo Qu¹, Yunsong Liu¹, Jing Ye¹, Di Guo², Zhifang Zhan¹, and Zhong Chen^1
1Department of Electronic Science, Xiamen University, Xiamen, Fujian, China, ²School of Computer and Information Engineering, Xiamen University of Technology, Xiamen, Fujian, China

Compressed sensing magnetic resonance imaging (CS-MRI) is to reconstruct MR images from undersampled k-space data by enforcing the sparsity of MR images. Patch-based nonlocal operator (PANO) is proposed as a linear operator to exploit the nonlocal self-similarity of MR images to further sparsify them. However, the original PANO is a frame and its numerical algorithm for CS-MRI problem is solved by the alternating direction minimization with continuation (ADMC). These two aspects lead the reconstruction to be time consuming. In this work, we first convert the PANO into a tight frame, and then applied the alternating direction method of multipliers (ADMM) algorithm to accelerate the image reconstruction. The empirical convergence demostrates that the new approach significantly accelerate the image reconstruction in compressed sensing MRI and can accomplish the reconstruction of one 256256 within several seconds.

Nishant Zachariah¹, Johannes M Flake², Qiu Wang³, Boris Mailhe³, Justin Romberg¹, Xiaoping Hu⁴, and Mariappan Nadar^3
1Department of Electrical and Computer Engineering, Georgia Institute of Technoloy, Atlanta, GA, United States, ²Department of Mathematics, Rutgers University, New Brunswick, NJ, United States, ³Imaging and Computer Vision, Siemens Corporate Technology, Princeton, NJ, United States, ⁴Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, GA, United States

Increasing MR image resolution, decreasing MR instrumentation noise and reconstructing high quality MR images from under sampled measurements are open challenges. In this paper we tackle these three problems under a novel non convex framework. We show that our method out performs state of the art techniques (quantitatively and qualitatively) for image super-resolution, denoising and under sampled reconstruction. In addition, we are able to recover regions of clinical interest with greatest fidelity thereby substantially aiding the clinical diagnostic process. Our powerful generic framework lends itself to tackling additional future applications such as image in-painting and blind de-convolution.

Matthew J. Muckley¹, Douglas C. Noll¹, and Jeffrey A. Fessler^2
1Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States, ²Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, United States

MRI scan times can be accelerated by combining parallel MRI with sparse models. These models give rise to optimization problems that are traditionally minimized with variable splitting algorithms that require tuning of penalty parameters. We review a new algorithm, BARISTA, that circumvents penalty parameter tuning while preserving convergence speed. We then propose a new optimized momentum update term for BARISTA that gives a theoretically-predicted factor of 2 increase in convergence speed of the cost function, terming the new algorithm OMBARISTA. Our optimization experiments agreed with the theory predictions, and we propose using OMBARISTA in place of BARISTA in general settings.

Rizwan Ahmad¹, Philip Schniter¹, and Orlando P. Simonetti^2
1Electrical and Computer Engineering, The Ohio State University, Columbus, Ohio, United States, ²Internal Medicine and Radiology, The Ohio State University, Columbus, Ohio, United States

Redundant dictionaries are routinely used to exploit rich structure in MR images. When using a redundant dictionary, however, the level of sparsity may vary across different groups of atoms, i.e., across “subdictionaries.” In this work, we propose a method, called Sparsity Adaptive Compressive Recovery (SCoRe), that adapts to the inherent level of sparsity in each subdictionary. Moreover, the proposed adaptation is data-driven and does not introduce any tuning parameters. For validation, results from digital phantom and real-time cine are presented.

Zongying Lai^1,2, Yunsong Liu¹, Di Guo³, Jing Ye¹, Zhifang Zhan¹, Zhong Chen¹, and Xiaobo Qu^1
1Department of Electronic Science, Xiamen University, Xiamen, Fujian, China, ²Department of Communication Engineering, Xiamen University, Fujian, China,³School of Computer and Information Engineering, Xiamen University of Technology, Xiamen, Fujian, China

Compressed sensing MRI can speed up imaging by undersampling k-space data. However, the sparse representation of magnetic resonance images affects the quality of reconstructed images. In this work, a graph-based compressed sensing MRI image reconstruction method is proposed. This method views an image patch as a vertex on graph and reorders the pixel to be smooth by traveling this graph with shortest path. Image reconstruciong from compressively sampled data shows that the proposed reconstruction method outperforms conventional wavelets in terms of visual quality and evaluation criteria.

Zechen Zhou¹, Niranjan Balu², Rui Li¹, Jinnan Wang^2,3, and Chun Yuan^1,2
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, China, ²Vascular Imaging Lab, Department of Radiology, University of Washington, Seattle, WA, United States, ³Philips Research North America, Briarcliff Manor, NY, United States

Parallel Imaging (PI) and Compressed Sensing (CS) enable accelerated MR imaging. However, the actual PI-CS reconstruction performance is usually limited by noise amplification and image boundary/structure blurring particularly at high reduction factor. In this work, a Gaussian Mixture Model (GMM) was optimized to promote structured sparsity and it was further merged into the SPIRiT framework as a regularization constraint. The proposed algorithm has demonstrated its improved performance for image boundary and detail structure preservation in accelerated 3D high resolution brain imaging.

Gabriel Ramos-Llordén¹, Hilde Segers¹, Willem Jan Palenstijn¹, Arnold J. den Dekker^1,2, and Jan Sijbers^1
1iMinds Vision-Lab, University of Antwerp, Antwerp, Antwerp, Belgium, ²Delft Center for Systems and Control, Delft University of Technology, Delft, Netherlands

In MRI reconstruction, undersampled data sets lead to ill-posed reconstruction problems. To regularize these problems, prior knowledge is commonly exploited. In this work, we introduce a new type of prior knowledge, partial discreteness, where part of the image is assumed to be homogeneous and can be well represented by a constant magnitude. We introduce this prior in the common algebraic reconstruction problem and propose an iterative algorithm to approximately solve it. It combines a penalized least squares reconstruction with an internal Bayesian segmentation. Results with synthetic data demonstrate that more detailedly restored images are obtained when partial discreteness is exploited

Andres Saucedo^1,2, Stamatios Lefkimmiatis³, Stanley Osher³, and Kyunghyun Sung^1,2
1Department of Radiological Sciences, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, California, United States, ²Biomedical Physics Interdepartmental Graduate Program, University of California Los Angeles, Los Angeles, California, United States, ³Department of Mathematics, University of California Los Angeles, Los Angeles, California, United States

This study introduces a novel constrained reconstruction technique that exploits both the local correlation of image data across multiple coils and the inherent non-local self-similarity property of images. Our approach is based within a non-local total variation regularization framework. The proposed method is applicable to both compressed sensing and parallel imaging, and demonstrates substantial advantages with regard to high levels of noise.

Ryan Wen Liu¹, Lin Shi², Simon C.H. Yu¹, and Defeng Wang^1,3
1Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong, ²Department of Medicine and Therapeutics, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong, ³Department of Biomedical Engineering and Shun Hing Institute of Advanced Engineering, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong

In this study, we propose to combine L₀ regularized tree-structured wavelet sparsity (TsWS) and second-order total generalized variation (TGV²) to reconstruct MR image from highly undersampled k-space data. In particular, the L₀ regularized TsWS could better represent the measure of sparseness in wavelet domain. TGV² is capable of maintaining trade-offs between artefact suppression and tissue feature preservation. To achieve solution stability, the corresponding minimization problem is decomposed into several simpler subproblems. Each of these subproblems has a closed-form solution or can be efficiently solved using existing optimization algorithms. Experimental results have demonstrated the superior performance of our proposed method.

Johannes F. M. Schmidt¹ and Sebastian Kozerke^1,2
1Institute for Biomedical Engineering, University and ETH Zurich, Zurich, Switzerland, ²Division of Imaging Sciences and Biomedical Engineering, King's College London, United Kingdom

A statistical approach to combine multiple denoising algorithms in MR image reconstruction to suppress reconstruction artifacts.

Xiaohui Liu¹, Jinhong Huang¹, Wufan Chen¹, and Yanqiu Feng^1
1Guangdong Provincial Key Laborary of Medical Image Processing, School of Biomedical Engineering, Southern Medical University, Guangzhou, Guangdong, China

Undersampled MRI reconstruction techniques based on Compressed Sensing (CS) exploiting sparsity which is implicit in MR images can provide significant help in reducing the scan time during clinical period, but remains challenging due to the requirement of high reconstruction accuracy. A novel algorithm is developed and tested in vivo for solving the MR image reconstruction problem due to Nesterov¡¯s smoothing scheme and convex conic optimization.

MingJian Hong¹, MengRan Lin¹, Feng Liu², and YongXin Ge^1
1ChongQing University, ChongQing, ChongQing, China, ²ITEE, The University of Queensland, QLD, Australia

This work presented a new model by enforcing both local and global sparsity, which captures both the patch-level and global sparse structures of the anatomical images. Using a model split approach, the image reconstruction quality can be iteratively further improved. Our simulation results demonstrate that, the proposed method outperform those existing methods using only the patch-level or global sparse structure.

Yunsong Liu¹, Jian-Feng Cai², Zhifang Zhan¹, Di Guo³, Jing Ye¹, Zhong Chen¹, and Xiaobo Qu^1
1Department of Electronic Science, Xiamen University, Xiamen, Fujian, China, ²Department of Mathematics, University of Iowa, Iowa City, Iowa, United States, ³School of Computer and Information Engineering, Xiamen University of Technology, Xiamen, Fujian, China

Compressed sensing (CS) has shown to be promising to accelerate magnetic resonance imaging (MRI). There are two different sparse models in CS-MRI: analysis and synthesis models with different assumptions and performance when a redundant tight frame is used. A new balance model is introduced into CS-MRI that can achieve the solutions of the analysis model, synthesis model and some in between by tuning the balancing parameter. It is found in this work that the typical balance model has a comparable performance with the analysis model in CS-MRI. Both of them achieve lower reconstructed errors than the synthesis model no matter what value the balancing parameter is. These observations are consistent for different tight frames used CS-MRI.

Florian Knoll^1,2, Martin Holler³, Thomas Koesters^1,2, and Daniel K Sodickson^1,2
1Center for Advanced Imaging Innovation and Research (CAI2R), NYU School of Medicine, New York, NY, United States, ²Bernard and Irene Schwartz Center for Biomedical Imaging, Department of Radiology, NYU School of Medicine, New York, New York, United States, ³Department of Mathematics and Scientific Computing, University of Graz, Graz, Austria

It was recently shown that simultaneously acquired data from state-of-the-art MR-PET systems can be reconstructed simultaneously using the concept of joint sparsity, yielding benefits for both MR and PET reconstructions. In this work we propose a new dedicated regularization functional for multi-modality imaging that exploits common structures of the MR and PET images. The two modalities are treated as single multi-channel images and an extension of the second order Total Generalized Variation functional for vector valued data is used as a dedicated multi-modality sparsifying transform.

Chenguang Peng¹, Rong Guo¹, Yicheng Chen¹, Yingmao Chen², Quanzheng Li³, Georges El Fakhr³, and Kui Ying^1
1Key Laboratory of Particle and Radiation Imaging, Ministry of Education, Department of Engineering, Beijing, China, ²Department of Nuclear Medicine, The general hospital of Chinese People's Liberation, Beijing, China, Beijing, China, ³Department of Radiology, Division of Nuclear Medicine and Molecular Imaging, Harvard Medical School, Boston, United States

PET is a practical medical imaging technique for brain function diagnosis. However, the low spatial resolution limits the use of PET in neurology and disease like Alzheimer's disease. With the help of MRI-PET, people can use high resolution MRI to provide anatomical information to correct partial volume effect of PET image which is a great cause for low resolution. Nevertheless, traditional partial volume effect correction method requires an accurate MRI segmentation and PVE model estimation which are not usually applicable. In this work, we proposed a method that is insensitive to PVE model estimation error and segmentation error.

Mathias Lukas¹, Anne Kluge², Jorge Cabello¹, Christine Preibisch^2,3, and Stephan Nekolla^1
1Department of Nuclear Medicine, Klinikum rechts der Isar, TU München, Munich, Germany, ²Department of Neuroradiology, Klinikum rechts der Isar, TU München, Munich, Germany, ³Department of Neurology, Klinikum rechts der Isar, TU München, Munich, Germany

Attenuation correction (AC) in quantitative PET/MR is affected by SNR and CNR of underlying MR sequences. In this work, the quality of MR data currently used for attenuation correction in PET (UTE, DIXON, MPRAGE) was observed in-vivo under changing clinical conditions over 3 months to investigate the reliability and robustness for in-house established MR based AC methods. In spite of its semi quantitative character, all sequences were found to be very invariant in SNR and CNR and can be used without any concerns.

Lishui Cheng¹, Sangtae Ahn¹, Dattesh Shanbhag², Florian Wiesinger³, Sandeep Kaushik², and Ravindra Manjeshwar^1
1GE Global Research, Niskayuna, NY, United States, ²GE Global Research, Bangalore, India, ³GE Global Research, Munich, Germany

Attenuation correction is critical to accurate PET quantitation. In PET/MR, MR-based attenuation correction (MR-AC) has challenges in bone, air, lung and implant regions. To address the problem, we combined 1) a segmentation-based MR-AC method, which works well in soft-tissue regions, and 2) a selective-update joint estimation approach, which reconstructs both attenuation and activity from PET emission data, to resolve the attenuation in the challenging regions. The method was evaluated on clinical data from a PET/MR scanner with TOF information and it was demonstrated that the method can distinguish between abdominal air and spinal implant/bone regions, otherwise hidden in MR.